implementation of a Probabilistic U-net on connectomic datasets. The model produces a (specified) number of predicted plausible segmentation samples for a given input volume.

Dependencies:

• gunpowder: required for building the data-loading pipeline.
• skelerator: required for generating the neural toy data used for testing the implementation
• nvidia-docker (optional): required for training the model in a docker container that leverages NVIDIA GPUs

To generate a set of data:

1. Go into data/, run mkdir datasets
2. Run python generate_full_samples n, where n is the number of data samples.

Training/Prediction:

• all scripts (.sh files) must be run in the base project directory:

To train a setup:

1. Run ./create.sh setup_X, where X is the setup name/number. This creates the relevant sub directories for the setup in train/, log/ and snapshots/
2. If wanting to use a previous result, copy the 5 python files from it into your new setup
3. Run ./mknet_train.sh setup_X to generate the training meta and config files
4. Run ./train.sh setup_X n where n is the number of iterations to train the network

To predict a setup:

1. Run ./mknet_predict setup_X to generate the predict meta and config files
2. Run ./predict.sh setup_X c n where c is the checkpoint number and n is the number of predictions